Machine Unlearning on RETURN
80Forget AccuracyOriginal
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| OriginalBackbone=DeepSeek-R1-Distill-Llama-8B2026.04 | 80 | 79.45 | 79.07 | |
| ReasonedIDKBackbone=DeepSeek-R1-Distill-Llama-8B2026.04 | 55.03 | 52.6 | 49.63 | |
| GDBackbone=DeepSeek-R1-Distill-Llama-8B2026.04 | 43.41 | 50.84 | 45.18 | |
| NPOBackbone=DeepSeek-R1-Distill-Llama-8B2026.04 | 37.21 | — | 22.96 | |
| R2MUBackbone=DeepSeek-R1-Distill-Llama-8B2026.04 | 34.88 | 45.35 | 70.37 | |
| CiPOBackbone=DeepSeek-R1-Distill-Llama-8B2026.04 | 31.78 | 44.46 | 81.48 | |
| GABackbone=DeepSeek-R1-Distill-Llama-8B2026.04 | 28.88 | — | 37.2 | |
| DirectIDKBackbone=DeepSeek-R1-Distill-Llama-8B2026.04 | 26.35 | 24.93 | 39.26 | |
| AnswerIDKBackbone=DeepSeek-R1-Distill-Llama-8B2026.04 | 5.42 | 63.54 | 10.37 |